What is geometric memory in AI video, and how does it keep characters and clothing consistent?

Geometric memory in AI video is a pre-built visual record — unified character-garment sheets showing each specific garment on each specific character — that teaches the generation system how a fabric interacts with that character's facial and body geometry. Built before any video generation, it keeps the same face and the same clothing behavior across every shot.

Geometric memory is built, not prompted: before generating a single video clip, you produce unified character-garment sheets — reference images that show each garment on each specific character from multiple angles — and lock them as the project's source of truth for how that fabric sits on that body. invideo is an agentic video creation tool where you can run this as a crew: assign a model designer sub-agent to produce the sheets while a creative producer agent and cinematic director agent handle shot planning in parallel — one documented production used four specialized parallel agents this way.

Why it prevents drift. Character and clothing inconsistency happens when each shot is generated from a fresh text prompt with no persistent record of the character's geometry — faces morph, garment folds and textures shift shot to shot. With the sheets locked, the invideo agent's context window holds the entire project memory — clothes, fabric physics, camera rules, and every locked parameter — so you never re-describe the character or the garment per shot. One documented fashion production kept 5 distinct characters visually consistent across different lighting conditions and heavy garments using exactly this setup.

How it's used during generation. Each video clip is generated with the character-garment sheet attached as a reference alongside the locked frame for that shot — Seedance 2.0 takes the character sheet and the static shot as dual references per clip, and in one documented test the same character held across both Kling and Seedance 2.0 in the same generation pass. All of these models run inside invideo, so the invideo agent routes each shot to the right one without you switching tools.

Complex motion runs on the same memory. For shots with complex movement, skip the storyboard entirely: feed the raw character sheet and garment blueprint, then describe the motion in natural language — the model calculates the fabric physics from the stored geometry instead of guessing from a flat frame.

When drift still slips through. If a character inconsistency appears after generation, issue a single text command — the invideo agent detects and regenerates every affected storyboard automatically, rather than requiring you to rebuild the character sheet and every frame from scratch.

Watch some of these to see what works for you:

By generating these unified character garment sheets before shooting, I build a total geometric memory. The agent one now knows exactly how that specific fabric interacts with the specific facial geometry of that model.

— invideo's creative team